Steering Reference Card

Summary — what this page covers A printable one-pager attendees keep. It answers "where does this instruction belong?" with the seven-methods decision table, the three decisions, a quick flowchart, and .NET copy-paste starters. Paste the finished content from the workshop's Steering Reference Card here; this stub gives the skeleton.

The one page worth keeping.

The question behind every instruction

Decide by four properties: when it loads · survives compaction · context cost · authority.

The seven methods

Method	When it loads	Survives compaction	Context cost	Use it for
CLAUDE.md (root)	Session start; re-read after compaction	Yes (memoized)	Always in context — keep < 200 lines	Facts & soft conventions Claude needs all the time
CLAUDE.md (subdir)	On-demand, when working in that folder	Partial (reloads when relevant)	Low — only when you're in that area	Folder-specific facts
Rules	Always (unscoped) or on file-match (path-scoped)	Yes	Low–medium	Path-scoped coding conventions
Skills	On `/name`, or auto-match by description	n/a (invoked on demand)	Low until triggered	A multi-step procedure (~30+ lines)
Subagents	When spawned	Isolated context; only the summary returns	Cheap to the parent — runs in its own window	Out-of-sight heavy work; a cost lever via `model:`
Hooks	On events (deterministic)	n/a (runs as code)	None — executes outside the model	"Every time X, do Y"; real guardrails (`PreToolUse` exit 2 blocks)
Output styles	Session config	n/a	Minimal	Changing how Claude formats responses
Append-system-prompt	Session start (flag)	Yes	Adds to every turn	One-off system-prompt additions

Three decisions you'll actually use

"Every time X, always do Y" → hook.
"Never do X" → not a CLAUDE.md line (hooks · permissions · managed settings).
A 30-line procedure → skill.

"Where does this go?" — quick flow

Walk it top to bottom; take the first match:

"Every time / never X" (deterministic)? → HOOK (PreToolUse exit 2 to block).
A multi-step procedure (~30+ lines)? → SKILL (a folder, triggered by its description).
A coding convention for a file area? → RULE (path-scoped glob).
An always-needed project fact? → CLAUDE.md (root = map; subdir = local facts).
Heavy / noisy work you want out of context? → SUBAGENT (isolated; summary returns).

.NET copy-paste starters

Path-scoped C# rule — .claude/rules/data-access.md

---
paths:
  - "BookTracker.Data/**/*.cs"
---
# Data-access conventions
- Parameterized queries only — never string-concatenated SQL.
- Async EF methods; thread the CancellationToken.
- Migrations live in BookTracker.Data.

Skill folder layout — .claude/skills/<name>/

.claude/skills/add-api-endpoint/
  SKILL.md          # name + model-facing description (the trigger) + procedure + Gotchas
  references/       # supporting docs the skill points Claude to
  scripts/          # helper scripts the skill can run

Subagent (.md + frontmatter) — .claude/agents/test-writer.md

---
name: test-writer
description: Generate xUnit tests for a class, matching the patterns in BookTracker.Tests.
model: claude-haiku-4-5
---
You write focused xUnit tests. Read the target class and existing tests first,
mirror the house style, and cover happy path + edge cases.

Guardrail hook registration — .claude/settings.json

{
  "hooks": {
    "PreToolUse": [
      { "matcher": "Bash", "hooks": [
        { "type": "command", "command": ".claude/hooks/block-destructive.sh" }
      ]}
    ],
    "PostToolUse": [
      { "matcher": "Edit|Write", "hooks": [
        { "type": "command", "command": "dotnet build" }
      ]}
    ]
  }
}

Gotchas worth remembering

Skills are folders, not files. Subagents are .md, not .yaml.
Path-scoped rules load on read, not create — make creation-time rules unscoped.
User-level ~/.claude/rules/ paths: is currently unreliable — use project-level.
A skill's description is a trigger written for the model; its Gotchas section is the high-value part.